Go to chat

Docs

Point a harness at MagmaRouter, understand what the caps do to your request, and find out which models can actually call tools.

Looking for your key...
Sign in

Every tile below copies setup text for that tool.

Power your agents with MagmaRouter

Set MagmaRouter as the model behind the agent you already run.

MagmaRouter skills

Give your agent MagmaRouter as a skill it can call for the work its own model turns down. Paste the instructions into the chat you have with your agent.

Endpoints

OpenAI formathttps://api.magmarouter.com/v1
Chat completions endpointhttps://api.magmarouter.com/v1/chat/completions
Anthropic formathttps://api.magmarouter.com

What MagmaRouter is

Get started

  • What MagmaRouter is

    MagmaRouter is an OpenAI- and Anthropic-compatible gateway for uncensored, abliterated and niche models, with per-key spend caps that hold and no prompt retention.

  • Quickstart: your first request

    Get a key, pick a model and make a billed request against MagmaRouter in about two minutes.

  • Quickstart: run Codex CLI on it

    Point Codex CLI at MagmaRouter through the Responses endpoint, with the same key, caps and receipts as every other route.

  • Quickstart: run Claude Code on it

    Point Claude Code at MagmaRouter with four environment variables. Verified end to end, tool calls included.

  • Quickstart: run the DeepSeek Harness on it

    Add MagmaRouter to the DeepSeek Harness as a custom provider, by form or by file, with the same key, caps and receipts as every other route.

The gateway

Spend

  • Keys

    Mint a key per user, per agent or per engagement, each with its own cap. Keys are stored as hashes and shown once.

  • Spend caps, and what happens at the edge

    Caps are enforced before the model runs: reserve the prompt cost, clamp max_tokens to what is left, settle on the real usage.

  • Billing

    Credit is prepaid and held in integer nano-dollars. Top ups happen inside the dashboard, not on a redirect.

Trust

  • Signed request receipts

    Every response carries a signed statement of which model served it and the hash of the canonical prompt. Verify it without handing back the prompt.

  • Retention

    Model, tokens, cost, status, timestamp. Not the prompt, not the completion, not a sample of either.

Reference

  • Endpoint reference

    Every public endpoint the gateway serves, what it takes and what it returns.

  • Errors

    Errors come back in the shape of the format you called, so your SDK parses them without special handling.

  • Known limits

    The things this gateway genuinely cannot do today, stated plainly so you find out here rather than halfway through an integration.